x86/aacpsdsp: add ff_ps_hybrid_analysis_ileave_sse

About 2x faster than the c version.

libavcodec/x86/aacpsdsp.asm

@@ -166,6 +166,112 @@ align 16
     jl .loop
     REP_RET
 
+;**********************************************************
+;void ps_hybrid_analysis_ileave_sse(float out[2][38][64],
+;                                   float (*in)[32][2],
+;                                   int i, int len)
+;**********************************************************
+INIT_XMM sse
+cglobal ps_hybrid_analysis_ileave, 3, 7, 5, out, in, i, len, in0, in1, tmp
+    movsxdifnidn        iq, id
+    mov               lend, 32 << 3
+    lea                inq, [inq+iq*4]
+    mov               tmpd, id
+    shl               tmpd, 8
+    add               outq, tmpq
+    mov               tmpd, 64
+    sub               tmpd, id
+    mov                 id, tmpd
+
+    test                id, 1
+    jne .loop4
+    test                id, 2
+    jne .loop8
+
+align 16
+.loop16:
+    mov               in0q, inq
+    mov               in1q, 38*64*4
+    add               in1q, in0q
+    mov               tmpd, lend
+
+.inner_loop16:
+    movaps              m0, [in0q]
+    movaps              m1, [in1q]
+    movaps              m2, [in0q+lenq]
+    movaps              m3, [in1q+lenq]
+    TRANSPOSE4x4PS 0, 1, 2, 3, 4
+    movaps          [outq], m0
+    movaps     [outq+lenq], m1
+    movaps   [outq+lenq*2], m2
+    movaps [outq+3*32*2*4], m3
+    lea               in0q, [in0q+lenq*2]
+    lea               in1q, [in1q+lenq*2]
+    add               outq, mmsize
+    sub               tmpd, mmsize
+    jg .inner_loop16
+    add                inq, 16
+    add               outq, 3*32*2*4
+    sub                 id, 4
+    jg .loop16
+    RET
+
+align 16
+.loop8:
+    mov               in0q, inq
+    mov               in1q, 38*64*4
+    add               in1q, in0q
+    mov               tmpd, lend
+
+.inner_loop8:
+    movlps              m0, [in0q]
+    movlps              m1, [in1q]
+    movhps              m0, [in0q+lenq]
+    movhps              m1, [in1q+lenq]
+    SBUTTERFLYPS 0, 1, 2
+    SBUTTERFLYPD 0, 1, 2
+    movaps          [outq], m0
+    movaps     [outq+lenq], m1
+    lea               in0q, [in0q+lenq*2]
+    lea               in1q, [in1q+lenq*2]
+    add               outq, mmsize
+    sub               tmpd, mmsize
+    jg .inner_loop8
+    add                inq, 8
+    add               outq, lenq
+    sub                 id, 2
+    jg .loop16
+    RET
+
+align 16
+.loop4:
+    mov               in0q, inq
+    mov               in1q, 38*64*4
+    add               in1q, in0q
+    mov               tmpd, lend
+
+.inner_loop4:
+    movss               m0, [in0q]
+    movss               m1, [in1q]
+    movss               m2, [in0q+lenq]
+    movss               m3, [in1q+lenq]
+    movlhps             m0, m1
+    movlhps             m2, m3
+    shufps              m0, m2, q2020
+    movaps          [outq], m0
+    lea               in0q, [in0q+lenq*2]
+    lea               in1q, [in1q+lenq*2]
+    add               outq, mmsize
+    sub               tmpd, mmsize
+    jg .inner_loop4
+    add                inq, 4
+    sub                 id, 1
+    test                id, 2
+    jne .loop8
+    cmp                 id, 4
+    jge .loop16
+    RET
+
 ;***********************************************************
 ;void ps_hybrid_synthesis_deint_sse4(float out[2][38][64],
 ;                                    float (*in)[32][2],

libavcodec/x86/aacpsdsp_init.c

@@ -44,6 +44,8 @@ void ff_ps_hybrid_synthesis_deint_sse(float out[2][38][64], float (*in)[32][2],
                                       int i, int len);
 void ff_ps_hybrid_synthesis_deint_sse4(float out[2][38][64], float (*in)[32][2],
                                        int i, int len);
+void ff_ps_hybrid_analysis_ileave_sse(float (*out)[32][2], float L[2][38][64],
+                                      int i, int len);
 
 av_cold void ff_psdsp_init_x86(PSDSPContext *s)
 {
@@ -52,6 +54,7 @@ av_cold void ff_psdsp_init_x86(PSDSPContext *s)
     if (EXTERNAL_SSE(cpu_flags)) {
         s->add_squares            = ff_ps_add_squares_sse;
         s->mul_pair_single        = ff_ps_mul_pair_single_sse;
+        s->hybrid_analysis_ileave = ff_ps_hybrid_analysis_ileave_sse;
         s->hybrid_synthesis_deint = ff_ps_hybrid_synthesis_deint_sse;
         s->hybrid_analysis        = ff_ps_hybrid_analysis_sse;
     }